Divergence Decoding: Inference-Time Unlearning via Auxiliary Models

• URL: http://arxiv.org/abs/2605.31293v1
• Date: Fri, 29 May 2026 13:29:01 GMT
• Title: Divergence Decoding: Inference-Time Unlearning via Auxiliary Models
• Abstract summary: Divergence Decoding (DD) is a mechanism that uses small auxiliary models to steer the logits of the Large Language Models away from specific data during inference.<n>We find the method decisively outperforms state-of-the-art (SOTA) baselines on unlearning benchmarks.
• Score: 2.1485350418225244
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large Language Models (LLMs) frequently memorize sensitive training data thereby creating significant privacy and copyright risks. Addressing these risks, i.e., removing such knowledge from an existing model checkpoint, has proven challenging as many unlearning methods lead to catastrophic utility loss or are ineffective for complex queries. We introduce Divergence Decoding (DD), a mechanism that uses small auxiliary models to steer the logits of the LLM away from specific data during inference. Training these models is straight forward, i.e., we use standard pre-training and fine-tuning setups. We find the method decisively outperforms state-of-the-art (SOTA) baselines on unlearning benchmarks across a variety of model and training dataset scales consistent with DD being an effective and inexpensive solution to unlearning. We then demonstrate that this steered distribution can be trivially distilled back into the base model. Since the method is generally applicable to any probabilistic model, we explore its efficacy outside of text generation and find evidence of generalization to the domain of images.

Related papers

• SPaRFT: Self-Paced Reinforcement Fine-Tuning for Large Language Models [51.74498855100541]
  Large language models (LLMs) have shown strong reasoning capabilities when fine-tuned with reinforcement learning (RL)<n>We propose textbfSPaRFT, a self-paced learning framework that enables efficient learning based on the capability of the model being trained.
  arXiv  Detail & Related papers  (2025-08-07T03:50:48Z)
• DUSE: A Data Expansion Framework for Low-resource Automatic Modulation Recognition based on Active Learning [17.651073556023167]
  We introduce a data expansion framework called Dynamic Uncertainty-driven Sample Expansion (DUSE)<n>DUSE uses an uncertainty scoring function to filter out useful samples from relevant AMR datasets.<n>Experiments demonstrate that DUSE consistently outperforms 8 coreset selection baselines in both class-balance and class-imbalance settings.
  arXiv  Detail & Related papers  (2025-07-16T08:09:41Z)
• Provable unlearning in topic modeling and downstream tasks [36.571324268874264]
  Provable guarantees for unlearning are often limited to supervised learning settings.<n>We provide the first theoretical guarantees for unlearning in the pre-training and fine-tuning paradigm.<n>We show that it is easier to unlearn pre-training data from models that have been fine-tuned to a particular task, and one can unlearn this data without modifying the base model.
  arXiv  Detail & Related papers  (2024-11-19T16:04:31Z)
• Distributionally robust self-supervised learning for tabular data [4.172010719137041]
  Learning robust representation in presence of error slices is challenging, due to high cardinality features and the complexity of constructing error sets.<n>Traditional robust representation learning methods are largely focused on improving worst group performance in supervised setting in computer vision.<n>Our approach utilizes an encoder-decoder model trained with Masked Language Modeling (MLM) loss to learn robust latent representations.
  arXiv  Detail & Related papers  (2024-10-11T04:23:56Z)
• Label-Free Model Failure Detection for Lidar-based Point Cloud Segmentation [15.779651238128562]
  We introduce label-free model failure detection for lidar-based point cloud segmentation.<n>We leverage different data characteristics by training a supervised and self-supervised stream for the same task to detect failure modes.<n>We perform a large-scale qualitative analysis and present LidarCODA, the first publicly available dataset with labeled anomalies in real-world lidar data.
  arXiv  Detail & Related papers  (2024-07-19T13:36:35Z)
• Can LLMs Separate Instructions From Data? And What Do We Even Mean By That? [60.50127555651554]
  Large Language Models (LLMs) show impressive results in numerous practical applications, but they lack essential safety features.<n>This makes them vulnerable to manipulations such as indirect prompt injections and generally unsuitable for safety-critical tasks.<n>We introduce a formal measure for instruction-data separation and an empirical variant that is calculable from a model's outputs.
  arXiv  Detail & Related papers  (2024-03-11T15:48:56Z)
• An Information Theoretic Approach to Machine Unlearning [43.423418819707784]
  To comply with AI and data regulations, the need to forget private or copyrighted information from trained machine learning models is increasingly important.<n>In this work, we address the zero-shot unlearning scenario, whereby an unlearning algorithm must be able to remove data given only a trained model and the data to be forgotten.<n>We derive a simple but principled zero-shot unlearning method based on the geometry of the model.
  arXiv  Detail & Related papers  (2024-02-02T13:33:30Z)
• Zero-shot Retrieval: Augmenting Pre-trained Models with Search Engines [83.65380507372483]
  Large pre-trained models can dramatically reduce the amount of task-specific data required to solve a problem, but they often fail to capture domain-specific nuances out of the box. This paper shows how to leverage recent advances in NLP and multi-modal learning to augment a pre-trained model with search engine retrieval.
  arXiv  Detail & Related papers  (2023-11-29T05:33:28Z)
• In-Context Unlearning: Language Models as Few Shot Unlearners [27.962361828354716]
  We propose a new class of unlearning methods for Large Language Models (LLMs) This method unlearns instances from the model by simply providing specific kinds of inputs in context, without the need to update model parameters. Our experimental results demonstrate that in-context unlearning performs on par with, or in some cases outperforms other state-of-the-art methods that require access to model parameters.
  arXiv  Detail & Related papers  (2023-10-11T15:19:31Z)
• Universal Domain Adaptation from Foundation Models: A Baseline Study [58.51162198585434]
  We make empirical studies of state-of-the-art UniDA methods using foundation models. We introduce textitCLIP distillation, a parameter-free method specifically designed to distill target knowledge from CLIP models. Although simple, our method outperforms previous approaches in most benchmark tasks.
  arXiv  Detail & Related papers  (2023-05-18T16:28:29Z)
• CAFA: Class-Aware Feature Alignment for Test-Time Adaptation [50.26963784271912]
  Test-time adaptation (TTA) aims to address this challenge by adapting a model to unlabeled data at test time. We propose a simple yet effective feature alignment loss, termed as Class-Aware Feature Alignment (CAFA), which simultaneously encourages a model to learn target representations in a class-discriminative manner.
  arXiv  Detail & Related papers  (2022-06-01T03:02:07Z)
• ALT-MAS: A Data-Efficient Framework for Active Testing of Machine Learning Algorithms [58.684954492439424]
  We propose a novel framework to efficiently test a machine learning model using only a small amount of labeled test data. The idea is to estimate the metrics of interest for a model-under-test using Bayesian neural network (BNN)
  arXiv  Detail & Related papers  (2021-04-11T12:14:04Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.